diff --git a/crates/sylpheed-formats/examples/dialog_pair_37.rs b/crates/sylpheed-formats/examples/dialog_pair_37.rs
new file mode 100644
index 00000000..cc821037
--- /dev/null
+++ b/crates/sylpheed-formats/examples/dialog_pair_37.rs
@@ -0,0 +1,58 @@
+//! Are the 37 equal-button-count `GP_DIALOG` pairs language pairs, or not?
+//!
+//! 26 of 65 adjacent pairs differ in BUTTON COUNT — two languages cannot, so those
+//! are unrelated dialogs. For the rest my language reading was left UNSUPPORTED
+//! rather than refuted, and both agents observed that nothing rewards closing it.
+//!
+//! A language pair must share its BUTTON NAMES and ROWS exactly (a locale changes
+//! glyphs, not layout) and differ only elsewhere. An unrelated pair will differ in
+//! button names or rows too.
+//!
+//! CONTROL: entries 2/3 (identical element sets) must come out as "buttons match".
+//!
+//!     cargo run -p sylpheed-formats --example dialog_pair_37
+use sylpheed_formats::{pak::PakArchive, ui_layout};
+use std::path::PathBuf;
+
+fn sig(ar: &PakArchive, i: usize) -> Option<(Vec<(String, i32)>, usize)> {
+    let by = ar.read(&ar.entries()[i]).ok()?;
+    let b = ui_layout::parse_build(&by)?;
+    let mut v: Vec<(String, i32)> = b
+        .elements
+        .iter()
+        .filter(|e| e.name.contains("btn"))
+        .map(|e| (e.name.clone(), e.rest().map(|k| k.y).unwrap_or(e.pivot_y as i32)))
+        .collect();
+    v.sort();
+    let n = v.len();
+    Some((v, n))
+}
+
+fn main() {
+    let root = PathBuf::from(std::env::var("SYLPHEED_DISC").expect("SYLPHEED_DISC"));
+    let ar = PakArchive::open(root.join("dat/GP_DIALOG.pak")).expect("GP_DIALOG");
+    let n = ar.entries().len();
+    let (mut match_btn, mut differ_btn, mut ctrl) = (0, 0, false);
+    let mut examples = 0;
+    for k in (0..n).step_by(2) {
+        let (Some((a, na)), Some((b, nb))) = (sig(&ar, k), sig(&ar, k + 1)) else { continue };
+        if na != nb {
+            continue; // the 26 already settled
+        }
+        if a == b {
+            match_btn += 1;
+            if k == 2 { ctrl = true }
+        } else {
+            differ_btn += 1;
+            if examples < 5 {
+                println!("   entries {k:3}/{:<3} buttons DIFFER", k + 1);
+                println!("        {:?}", a.iter().map(|x| &x.0).collect::<Vec<_>>());
+                println!("        {:?}", b.iter().map(|x| &x.0).collect::<Vec<_>>());
+                examples += 1;
+            }
+        }
+    }
+    println!("\nequal-button-count pairs whose button NAMES+ROWS match : {match_btn}");
+    println!("equal-button-count pairs whose buttons DIFFER          : {differ_btn}");
+    println!("control (entries 2/3 counted as matching): {}", if ctrl { "PASSED" } else { "FAILED" });
+}
diff --git a/crates/sylpheed-formats/examples/dialog_pair_diffs.rs b/crates/sylpheed-formats/examples/dialog_pair_diffs.rs
new file mode 100644
index 00000000..10027fe8
--- /dev/null
+++ b/crates/sylpheed-formats/examples/dialog_pair_diffs.rs
@@ -0,0 +1,43 @@
+//! What differs between equal-button-count `GP_DIALOG` adjacent pairs?
+//!
+//! All 39 share button names and rows. That is consistent with a LANGUAGE PAIR and
+//! equally with TWO DIALOGS SHARING A BUTTON TEMPLATE (two yes/no boxes differing
+//! only in their message sprite). The difference is in what else differs: a
+//! language pair should differ in the SAME slots with locale-marked names.
+//!
+//!     cargo run -p sylpheed-formats --example dialog_pair_diffs
+use sylpheed_formats::{pak::PakArchive, ui_layout};
+use std::collections::BTreeSet;
+use std::path::PathBuf;
+
+fn names(ar: &PakArchive, i: usize) -> Option<BTreeSet<String>> {
+    let by = ar.read(&ar.entries()[i]).ok()?;
+    let b = ui_layout::parse_build(&by)?;
+    Some(b.elements.iter().map(|e| e.name.clone()).collect())
+}
+
+fn main() {
+    let root = PathBuf::from(std::env::var("SYLPHEED_DISC").expect("SYLPHEED_DISC"));
+    let ar = PakArchive::open(root.join("dat/GP_DIALOG.pak")).expect("GP_DIALOG");
+    let n = ar.entries().len();
+    let mut shown = 0;
+    for k in (0..n).step_by(2) {
+        let (Some(a), Some(b)) = (names(&ar, k), names(&ar, k + 1)) else { continue };
+        if a == b {
+            continue;
+        }
+        let oa: Vec<&String> = a.difference(&b).collect();
+        let ob: Vec<&String> = b.difference(&a).collect();
+        // only the equal-button-count ones
+        let ba = a.iter().filter(|x| x.contains("btn")).count();
+        let bb = b.iter().filter(|x| x.contains("btn")).count();
+        if ba != bb {
+            continue;
+        }
+        if shown < 6 {
+            println!("entries {k:3}/{:<3}  shared {}  only-in-{k}: {:?}  only-in-{}: {:?}",
+                     k + 1, a.intersection(&b).count(), oa, k + 1, ob);
+            shown += 1;
+        }
+    }
+}
diff --git a/docs/re/data/difficulty-is-a-dialog.txt b/docs/re/data/difficulty-is-a-dialog.txt
index 063d9273..5ad81ba7 100644
--- a/docs/re/data/difficulty-is-a-dialog.txt
+++ b/docs/re/data/difficulty-is-a-dialog.txt
@@ -187,3 +187,45 @@
 # rest on the pairing -- it rests on the unique four-button geometry (0 rivals
 # disc-wide) plus the oracle capture -- but the pairing was stated as fact and was
 # not one.
+
+################################################################################
+# ❌ AND THE 37 ARE NOW REFUTED TOO, not merely unsupported. 2026-08-31.
+#
+# Both agents recorded that the language reading survived for the 37 pairs that
+# differ WITHOUT a button-count mismatch, and both noted that nothing rewarded
+# closing it. It cost one scan.
+#
+# STEP 1: all 39 equal-button-count pairs share their button NAMES AND ROWS
+# exactly (examples/dialog_pair_37.rs; control: entries 2/3 counted as matching).
+# ⚠️ That does NOT settle it -- two different dialogs sharing a button template,
+# like two yes/no boxes, look identical by that test.
+#
+# STEP 2: look at what actually differs (examples/dialog_pair_diffs.rs):
+#
+#   entries 10/11   pzstg10_01..20 + _eff   vs   pzstg02_01..16 + _eff
+#   entries 12/13   pzstg11_*               vs   pzstg03_*
+#   entries 14/15   pzstg12_*               vs   pzstg13_*
+#   entries 16/17   pzstg04_*               vs   pzstg14_*
+#   entries 18/19   pzstg05_*               vs   pzstg15_*
+#   entries 20/21   pzstg06_*               vs   pzstg16_*
+#
+# ✅ THESE ARE DIFFERENT STAGES. They are the sixteen DLG_STAGE_TITLE01..16
+# dialogs from the table above, and an adjacent pair carries TWO DIFFERENT STAGES,
+# not two languages of one. The sprite COUNTS differ too (stg10 has 20 title
+# sprites, stg02 has 16), which is a different amount of text, not a translation
+# of the same text.
+#
+# => The language reading is refuted for the 37 as well. The whole 63 is now
+#    explained by one fact: **adjacent GP_DIALOG entries are unrelated dialogs.**
+#    No residue, no special case.
+#
+# 📌 THE 2 REMAINING PAIRS (0/1 and 2/3) are still the only ones with identical
+# element sets, and still unexplained -- but they are no longer anomalous against
+# a hypothesis, because the hypothesis is gone. Whether they are a language pair,
+# a duplicate, or two dialogs sharing every element is untested.
+#
+# ⚠️ WORTH RECORDING ABOUT THE PROCESS, not the disc: this question was left open
+# by both agents with the explicit observation that "nothing rewards closing it".
+# It cost two scans and went AGAINST the reading I had offered. A bound nobody is
+# incentivised to test is exactly where a convenient claim survives -- and the
+# next reader cannot tell whether a bound was respected or merely never revisited.